Entire-barn dairy cow tracking framework for multi-camera systems
Yota Yamamoto, Kazuhiro Akizawa, S. Aou, Yukinobu Taniguchi
Abstract
• This is the first attempt to track dairy cows across an entire barn by using multiple cameras. • The proposed method tracks cows using location information through advanced projective transformation and integration algorithms. • The tracking performance improved by 50% compared to methods relying on image features. • The proposed method uses location data, making it robust to cow appearance changes and lens distortion effects. The early detection of abnormalities, rapid diagnosis and prevention of diseases, and maintenance of desirable pregnancy schedules are crucial to ensure stable milk production. Non-intrusive and non-contact health management techniques based on image analysis have thus attracted much attention. The health status of dairy cows can be estimated by maintaining individual tracking records, such as the level of walking and the number of visits to feeding and drinking stations. In this paper, we propose a location-based tracking method for multiple dairy cows that can, for the first time, use multi-camera systems to consistently track dairy cows throughout the entire barn. Unlike existing methods that rely on image features, the proposed method uses only location information, making it robust to cow appearance features and the view warping caused by camera lens distortion. Our experiments on video datasets of moving and close packed cows show that our proposed method achieves an accuracy of about 90% for MOTA and about 80% for IDF1. The experiments demonstrate that our method consistently outperforms existing image feature-based methods.